Yao Liu

HydraHead achieves a remarkable 69% performance boost in long-context tasks by intelligently hybridizing attention mechanisms at the head level, challenging traditional layer-wise approaches.

Charge-based ranking can recover 99.2% of the performance gap in legal case retrieval, challenging the reliance on complex models for relevance assessment.

FAiT outperforms traditional Transformers by capturing evolving spectral characteristics in time series data, enabling sharper forecasts of transient signals.

Expert revisions of legal text can be leveraged as counterfactuals to significantly boost conflict classification accuracy.